Lablup said it plans to turn concerns about excess investment in AI infrastructure into a growth opportunity ahead of its KOSDAQ listing on November 3. CEO Shin Jeong-kyu said at an initial public offering press conference in Seoul on the 19th that software which improves use of existing infrastructure, without owning hardware, can cushion fluctuations in hardware investment. Asked about concerns over AI debt linked to CoreWeave, he said software’s role in reducing volatility caused by different rates of progress in hardware and software would become more important.

Founded in 2015, Lablup makes software for operating AI infrastructure. Its main platform, Backend.AI, supports accelerator management, including GPUs and NPUs, as well as model development and training, inference and serving, and agent and token operations. The platform combines GPU partitioning and virtualization developed by the company, workload orchestration, support for different accelerators, storage virtualization and inference services. It supports more than 11 accelerator types, including Nvidia, AMD and Intel GPUs and domestically developed NPUs. Lablup is also developing partitioning and virtualization technology for next-generation NPUs from FuriosaAI and Rebellions.

Backend.AI operates at more than 120 sites in South Korea and overseas, including Samsung Electronics, Hyundai Mobis, KT and Shinhan Bank. The company says it has managed more than 2,000 GPUs at one site, increased customer GPU utilization from 30% to 85%, or about 2.8 times, and cut GPU budgets by 40%. In 2021, it became the first software partner in Asia-Pacific selected for Nvidia’s DGX-Ready program. For the government’s independent AI foundation model project, Lablup operates training infrastructure for the Upstage consortium and has supported training with a cumulative total of more than 2,500 Nvidia B200 and H100 GPUs. In the “AI for Everyone” project, it provides token generation and routing technology to the KT consortium. Shin said these projects do not directly generate monetary returns, but build experience managing hundreds or thousands of GPUs together and contribute to scaling up.

Lablup sells its software as an integrated platform through annual subscriptions, with license fees based on the types and quantities of GPUs customers own. Recurring revenue accounted for 94.7% of total revenue in 2025, and its average net revenue retention rate for the most recent two-year period was 131.9%. Annual revenue growth was close to 50% over the most recent three years. Shin said the software-focused business has maintained operating profitability and that operating profit is expected to increase further in 2027. The company plans to unveil Backend.AI:GO 2.0 this month, extending the platform from data centers to edge environments such as personal computers. It is also commercializing Continuum Router, which automatically distributes requests, and Continuum Hub, which manages token use and costs.

Shin said the AI market has moved through a research-centered “age of science” and a technology-optimization “age of engineering” and entered an “age of industry,” where large-scale capital investment matters. He said the company pursued a listing because reinvesting existing earnings alone would make it difficult to keep pace with industrial growth. About half of the offering proceeds will go toward next-generation AI hardware, including Nvidia Vera Rubin, and technology demonstrations. Lablup plans to use about 30% of equipment operating time for technology testing and the remaining 70% for token generation. The company aims to cover operating costs through token production and analyze customers’ break-even points.

The rest of the offering proceeds will fund overseas expansion. Lablup plans to resume operations at its U.S. subsidiary in earnest early next year and step up its push into Japan and Southeast Asia. It is negotiating with double-digit numbers of overseas neocloud companies; some talks have reached the quotation stage. The company is also working to expand local sales through global resellers. Shin said the KOSDAQ listing would support expansion from GPU-centered infrastructure management into AI model and token operations, as Lablup seeks to grow into a global AI infrastructure software company.

Lablup said it plans to turn concerns about excess investment in AI infrastructure into a growth opportunity ahead of its KOSDAQ listing on November 3. CEO Shin Jeong-kyu said at an initial public offering press conference in Seoul on the 19th that software which improves use of existing infrastructure, without owning hardware, can cushion fluctuations in hardware investment. Asked about concerns over AI debt linked to CoreWeave, he said software’s role in reducing volatility caused by different rates of progress in hardware and software would become more important.

Founded in 2015, Lablup makes software for operating AI infrastructure. Its main platform, Backend.AI, supports accelerator management, including GPUs and NPUs, as well as model development and training, inference and serving, and agent and token operations. The platform combines GPU partitioning and virtualization developed by the company, workload orchestration, support for different accelerators, storage virtualization and inference services. It supports more than 11 accelerator types, including Nvidia, AMD and Intel GPUs and domestically developed NPUs. Lablup is also developing partitioning and virtualization technology for next-generation NPUs from FuriosaAI and Rebellions.

Backend.AI operates at more than 120 sites in South Korea and overseas, including Samsung Electronics, Hyundai Mobis, KT and Shinhan Bank. The company says it has managed more than 2,000 GPUs at one site, increased customer GPU utilization from 30% to 85%, or about 2.8 times, and cut GPU budgets by 40%. In 2021, it became the first software partner in Asia-Pacific selected for Nvidia’s DGX-Ready program. For the government’s independent AI foundation model project, Lablup operates training infrastructure for the Upstage consortium and has supported training with a cumulative total of more than 2,500 Nvidia B200 and H100 GPUs. In the “AI for Everyone” project, it provides token generation and routing technology to the KT consortium. Shin said these projects do not directly generate monetary returns, but build experience managing hundreds or thousands of GPUs together and contribute to scaling up.

Lablup sells its software as an integrated platform through annual subscriptions, with license fees based on the types and quantities of GPUs customers own. Recurring revenue accounted for 94.7% of total revenue in 2025, and its average net revenue retention rate for the most recent two-year period was 131.9%. Annual revenue growth was close to 50% over the most recent three years. Shin said the software-focused business has maintained operating profitability and that operating profit is expected to increase further in 2027. The company plans to unveil Backend.AI:GO 2.0 this month, extending the platform from data centers to edge environments such as personal computers. It is also commercializing Continuum Router, which automatically distributes requests, and Continuum Hub, which manages token use and costs.

Shin said the AI market has moved through a research-centered “age of science” and a technology-optimization “age of engineering” and entered an “age of industry,” where large-scale capital investment matters. He said the company pursued a listing because reinvesting existing earnings alone would make it difficult to keep pace with industrial growth. About half of the offering proceeds will go toward next-generation AI hardware, including Nvidia Vera Rubin, and technology demonstrations. Lablup plans to use about 30% of equipment operating time for technology testing and the remaining 70% for token generation. The company aims to cover operating costs through token production and analyze customers’ break-even points.

The rest of the offering proceeds will fund overseas expansion. Lablup plans to resume operations at its U.S. subsidiary in earnest early next year and step up its push into Japan and Southeast Asia. It is negotiating with double-digit numbers of overseas neocloud companies; some talks have reached the quotation stage. The company is also working to expand local sales through global resellers. Shin said the KOSDAQ listing would support expansion from GPU-centered infrastructure management into AI model and token operations, as Lablup seeks to grow into a global AI infrastructure software company.